def maxUnpool2d(inputs, indices, ks):
if (len(inputs.shape) != 4):
raise Exception("input({})不符合[miniBatch*inChannels*H*W]的形状要求".format(
inputs.shape))
indices = dim.vector(indices)
if (inputs.shape != indices.shape):
raise Exception("input({})与indices({})的形状不一致".format(
inputs.shape, indices.shape))
a = []
ks = int(ks)
for i, channel in enumerate(inputs):
a.append([])
for j, kernel in enumerate(channel):
a[i].append([])
h = kernel.shape[0]
w = kernel.shape[1]
q = []
for k in range(h):
p = []
for l in range(w):
factor = dim.zeros([ks, ks])
r = math.floor(indices[i, j, k, l] / ks)
c = math.floor(indices[i, j, k, l] % ks)
factor[r, c] = 1
p.append(kernel[k, l] * factor)
q.append(dim.concat(p, 1))
a[i][j] = dim.concat(q, 0)
rst = dim.vector(a)
return rst
def maxUnpool1d(inputs, indices, ks):
if (len(inputs.shape) != 3):
raise Exception("input({})不符合[miniBatch*inChannels*W]的形状要求".format(
inputs.shape))
print("maxUnpool1d indices:", indices)
indices = dim.vector(indices)
if (inputs.shape != indices.shape):
raise Exception("input({})与indices({})的形状不一致".format(
inputs.shape, indices.shape))
a = []
ks = int(ks)
for i, channel in enumerate(inputs):
a.append([])
for j, kernel in enumerate(channel):
a[i].append([])
p = []
w = kernel.size
for k in range(w):
factor = dim.zeros(ks)
r = math.floor(indices[i, j, k] % ks)
factor[r] = 1
p.append(kernel[k] * factor)
a[i][j] = dim.concat(p, 0)
rst = dim.vector(a)
return rst
def avgPool2d(inputs, ks, padding=0):
if (len(inputs.shape) != 4):
raise Exception("input({})不符合[miniBatch*inChannels*H*W]的形状要求".format(
inputs.shape))
ks = int(ks)
a = []
for i, channel in enumerate(inputs):
a.append([])
for j, kernel in enumerate(channel):
a[i].append([])
kernel = kernel.pad(padding)
ih = kernel.shape[0]
iw = kernel.shape[1]
fh = ks
fw = ks
w = math.floor((iw - fw) / ks + 1)
h = math.floor((ih - fh) / ks + 1)
for k in range(h):
a[i][j].append([])
for l in range(w):
flip = kernel[k * ks:k * ks + fh, l * ks:l * ks + fw]
a[i][j][k].append(flip.avg())
rst = dim.vector(a)
rst = inputs.setGradFn(rst, "avgPool2d", left=inputs, right=ks)
return rst
def convTranspose1d(inputs, filters, stride=1, padding=0):
if (len(inputs.shape) != 3):
raise Exception("input({})不符合[miniBatch*inChannels*W]的形状要求".format(
inputs.shape))
if (len(filters.shape) != 3):
raise Exception("filter({})不符合[inChannels*outChannels*W]的形状要求".format(
filters.shape))
if (inputs.shape[1] != filters.shape[0]):
raise Exception("input({})与filter({})中channels数不一致".format(
inputs.shape, filters.shape))
#change channel
filters = dim.swapaxes(filters, 0, 1)
a = []
for i in range(inputs.shape[0]): #miniBatch
a.append([])
for j in range(filters.shape[0]): #outChannel
a[i].append([])
for k in range(filters.shape[1]): #kernel
bat = inputs[i, k].pad(padding)
kernel = filters[j, k].rot180()
iw = bat.size
fw = kernel.size
w = math.floor((iw - fw) / stride + 1)
for l in range(w):
value = (bat[l * stride:l * stride + fw] * (kernel)).sum()
if len(a[i][j]) <= l:
a[i][j].append(value)
else:
a[i][j][l] += value
rst = dim.vector(a)
return rst
def maxPool1d(inputs, ks, padding=0, includeIndices=False):
if (len(inputs.shape) != 3):
raise Exception("input({})不符合[miniBatch*inChannels*W]的形状要求".format(
inputs.shape))
ks = int(ks)
a = []
indices = []
for i, channel in enumerate(inputs):
a.append([])
indices.append([])
for j, kernel in enumerate(channel):
a[i].append([])
indices[i].append([])
kernel = kernel.pad(padding)
iw = kernel.size
fw = ks
w = math.floor((iw - fw) / ks + 1)
for k in range(w):
flip = kernel[k * ks:k * ks + fw]
a[i][j].append(flip.max())
indices[i][j].append(flip.argmax())
rst = dim.vector(a)
rst = inputs.setGradFn(rst,
"maxPool1d",
left=inputs,
right=ks,
args={"indices": indices})
if (includeIndices): return rst, indices
return rst
def softmax(x, axis=1):
x = dim.vector(x)
y = x.exp()
total = y.sum(axis, keepdims=True)
rst = y / total
if (axis == 1):
rst = x.setGradFn(rst, "softmax")
return rst
def crossEntropy(a, y):
a, y = dim.vector(a, y)
b = softmax(a, 1)
y_onehot = y.onehot(b.shape[1])
rst = y_onehot.mul(b.log()).sum(1).neg().mean()
rst = a.setGradFn(rst, "crossEntropy", right=y)
return rst
def softmaxDeri(x, a):
x, a = dim.vector(x, a)
argmax = a.argmax(1).value()
data = x.value
'''rst = Vector(data.map((y,i)=>{
return y.map((z,j)=>argmax[i]==j?z*(1-z):-z*y[argmax[i]])
}))
'''
return rst
def avgUnpool1d(inputs, ks):
if (len(inputs.shape) != 3):
raise Exception("input({})不符合[miniBatch*inChannels*W]的形状要求".format(
inputs.shape))
ks = int(ks)
a = []
factor = dim.fill(1 / ks, ks)
for i, channel in enumerate(inputs):
a.append([])
for j, kernel in enumerate(channel):
a[i].append(dim.kron(kernel, factor))
rst = dim.vector(a)
return rst
def conv2d(inputs, filters, stride=1, padding=0):
if (len(inputs.shape) != 4):
raise Exception("input({})不符合[miniBatch*inChannels*H*W]的形状要求".format(
inputs.shape))
if (len(filters.shape) != 4):
raise Exception(
"filter({})不符合[outChannels*inChannels*H*W]的形状要求".format(
filters.shape))
if (inputs.shape[1] != filters.shape[1]):
raise Exception("input({})与filter({})中channels数不一致".format(
inputs.shape, filters.shape))
a = []
for i in range(inputs.shape[0]):
a.append([])
for j in range(filters.shape[0]):
a[i].append([])
for k in range(filters.shape[1]):
bat = inputs[i, k].pad(padding)
kernel = filters[j, k]
ih = bat.shape[0]
iw = bat.shape[1]
fh = kernel.shape[0]
fw = kernel.shape[1]
w = math.floor((iw - fw) / stride + 1)
h = math.floor((ih - fh) / stride + 1)
for l in range(h):
if len(a[i][j]) <= l:
a[i][j].append([])
for m in range(w):
value = (bat[l * stride:l * stride + fh,
m * stride:m * stride + fw] *
(kernel)).sum().value()
if len(a[i][j][l]) <= m:
a[i][j][l].append(value)
else:
a[i][j][l][m] += value
rst = dim.vector(a)
#P1=(((In-1)*S+F) -(((In-F+2*P0)/S)+1))/2let In=input.shape[2]
In = inputs.shape[2]
F = filters.shape[2]
S = stride
P0 = padding
gradPadding = int((((In - 1) * S + F) - (((In - F + 2 * P0) / S) + 1)) / 2)
rst = inputs.setGradFn(rst,
"conv2d",
left=inputs,
right=filters,
args={"padding": gradPadding})
return rst
def convTranspose3d(inputs, filters, stride=1, padding=0):
#要实现还原运算，padding=((Out-1)*stride-Input+Filter)/2
if (len(inputs.shape) != 5):
raise Exception("input({})不符合[miniBatch*inChannels*D*H*W]的形状要求".format(
inputs.shape))
if (len(filters.shape) != 5):
raise Exception(
"filter({})不符合[outChannels*inChannels*D*H*W]的形状要求".format(
filters.shape))
if (input.shape[1] != filter.shape[0]):
raise Exception("input({})与filter({})中channels数不一致".format(
inputs.shape, filters.shape))
filters = dim.swapaxes(filters, 0, 1)
a = []
for i in range(inputs.shape[0]):
a.append([])
for j in range(filters.shape[0]):
a[i].append([])
for k in range(filters.shape[1]):
bat = inputs[i, k].pad(padding)
kernel = filters[j, k].rot180()
ideep = bat.shape[0]
ih = bat.shape[1]
iw = bat.shape[2]
fdeep = kernel.shape[0]
fh = kernel.shape[1]
fw = kernel.shape[2]
d = math.floor((ideep - fdeep) / stride + 1)
w = math.floor((iw - fw) / stride + 1)
h = math.floor((ih - fh) / stride + 1)
for l in range(d):
if len(a[i][j]) <= l: a[i][j].append([])
for m in range(h):
if len(a[i][j][l]) <= m: a[i][j][l].append([])
for n in range(w):
value = (bat[l * stride:l * stride + fdeep,
m * stride:m * stride + fh,
n * stride:n * stride + fw] *
(kernel)).sum().value()
if len(a[i][j][l][m]) <= n:
a[i][j][l][m].append(value)
else:
a[i][j][l][m][n] += value
return dim.vector(a)
def conv1d(inputs, filters, stride=1, padding=0):
if (len(inputs.shape) != 3):
raise Exception("input({})不符合[miniBatch*inChannels*W]的形状要求".format(
inputs.shape))
if (len(filters.shape) != 3):
raise Exception("filter({})不符合[outChannels*inChannels*W]的形状要求".format(
filters.shape))
if (inputs.shape[1] != filters.shape[1]):
raise Exception("input({})与filter({})中channels数不一致".format(
inputs.shape, filters.shape))
a = []
for i in range(inputs.shape[0]): #miniBatch
a.append([])
for j in range(filters.shape[0]): #outChannel
a[i].append([])
for k in range(filters.shape[1]): #inChannel
bat = inputs[i, k].pad(padding)
kernel = filters[j, k]
iw = bat.size
fw = kernel.size
w = math.floor((iw - fw) / stride + 1)
for l in range(w):
value = bat[l * stride:l * stride + fw].dot(kernel)
if len(a[i][j]) <= l:
a[i][j].append(value)
else:
a[i][j][l] += value
rst = dim.vector(a)
#P1=(((In-1)*S+F) -(((In-F+2*P0)/S)+1))/2let In=input.shape[2]
In = inputs.shape[2]
F = filters.shape[2]
S = stride
P0 = padding
gradPadding = math.floor(
((((In - 1) * S + F) - (((In - F + 2 * P0) / S) + 1)) / 2))
rst = inputs.setGradFn(rst,
"conv1d",
left=inputs,
right=filters,
args={"padding": gradPadding})
return rst
def __init__(self, data, name=None):
super(Constant, self).__init__()
if self.isNumber(data):
self.number = data
self.data = dim.vector(data)
elif (isinstance(data, dim.Vector) and data.size == 1):
self.number = data
self.data = data
elif (isinstance(data, dim.cl.Array) and data.size == 1):
self.number = data
self.data = data
else:
self.number = None
self.data = data
'''节省空间，但严重影响效率
md5=hashlib.md5()
md5.update(json.dumps(data.tolist()).encode())
hashData = md5.hexdigest()
try:
idx=list((x["hash"] for x in CONSTANT)).index(hashData)
obj = CONSTANT[idx]["object"]
self.name=obj.name
self.data=obj.data
except :
self.name = "const"+str(random.random())[-6:]
self.data = data
CONSTANT.append({"name":self.name,"hash":hashData,"object":self})
'''
if (name is None):
Constant.sequence += 1
self.name = "const" + str(Constant.sequence)
else:
self.name = name
#print(self.name,self.number,data,type(data))
self.type = "Constant"
self._expressionStr = self.name
def binaryCrossEntropy(a, y):
a, y = dim.vector(a, y)
return (y * a.log() + (1 - y) * (1 - a).log()).sum()
def tanh(x):
x = dim.vector(x)
return x.tanh()
def dropout(a, keep):
if (keep <= 0 or keep > 1): raise Exception("keep_prob参数必须属于(0,1]")
a = dim.vector(a)
arr = []
'''return new Vector(a.data.map((x,i)=>{
def sigmoid(x):
x = dim.vector(x)
return 1 / (1 + (-x).exp())
def crossEntropyDeri(a, y):
a, y = dim.vector(a, y)
b = softmax(a, 1)
y_onehot = dim.onehot(y, a.shape[1])
rst = b.sub(y_onehot).div(b.shape[0])
return rst
def softplus(x):
x = dim.vector(x)
return (x.exp() + 1).log()
def logcoshLoss(a, y):
a, y = dim.vector(a, y)
return y.sub(a).cosh().log().sum()
def get(self):
return dim.vector(cl.array.Array.get(self))